README.md

---
language:
- multilingual
tags:
- audio
- deepfake-detection
- speech
- synthetic-speech
- anti-spoofing
- acoustic-features
- biomechanics
- tts
- mlaad
task_categories:
- audio-classification
size_categories:
- 100K<n<1M
pretty_name: MLAAD_Audit
viewer: false
---

# MLAAD — SSA Acoustic Feature Audit

**Moonscape Software | Synthetic Speech Atlas**
*Research audit contribution to the MLAAD dataset team*

---

## Overview

This repository contains acoustic feature measurements extracted from the
**MLAAD (Multilingual Audio Anti-Spoofing Dataset)** corpus by the Moonscape
**Synthetic Speech Atlas (SSA)** pipeline.

**298,000 rows. 152 columns. One row per MLAAD clip.**

No audio files are included. Each row contains classical signal processing
and biomechanical acoustic features extracted from the original MLAAD audio,
plus Z-scores computed against the SSA biological baseline (VCTK anechoic
chamber + AMI conversational speech).

The SSA pipeline is physics-grounded rather than learned — every feature
measures a physical property of the vocal production system. This makes the
measurements interpretable and directly comparable across datasets without
retraining.

---

## Files

| File | Description |
|------|-------------|
| `mlaad_with_zscores_20260522.parquet` | 298,000 rows × 152 columns. Raw acoustic features + Z-scores. |
| `docs/ssa_DATA_DICTIONARY.txt` | Full column-by-column reference with null counts and baseline parameters. |
| `docs/ssa_METHODOLOGY.txt` | SSA six-pass pipeline documentation and key findings. |

---

## Column Structure

```
Cols   1-9    Provenance metadata (data_provider, contact, distribution, etc.)
Cols  10-19   Identity and clip metadata (file_id, language, tts_system, tier, etc.)
Cols  20-93   Raw acoustic features (74 columns)
Cols  94-152  Z-score columns, Z_ prefix (59 columns)
```

### Key Features

**Primary SSA detection signals:**

| Column | Description | Typical spoof Z |
|--------|-------------|-----------------|
| `bico_f0_f1` | F0-F1 bicoherence coupling. Primary detection signal. TTS destroys the nonlinear coupling between glottal source and vocal tract filter. | −0.3 to −1.5 |
| `spectral_tilt` | Log power spectrum slope. Only universally direction-consistent feature across all tested architectures. | +0.1 to +0.5 |
| `modgd_var` | Modified Group Delay variance. Most robust under adversarial conditions. | −0.2 to −0.8 |
| `nVIV` | Normalised Voiced Interval Variance. Rhythm metric. TTS collapses toward English stress-timed values regardless of target language. | varies by language |
| `iaif_residual_kurtosis` | Glottal residual kurtosis. Biological speech has leptokurtic GCI spikes; synthetic is near-Gaussian. | −0.2 to −0.5 |
| `f0_declination_slope` | F0 trajectory slope. TTS has no lung pressure model — near-zero or artificially reset. | +0.3 to +1.0 |

**Z-score interpretation:**
- `|Z| > 3.0` → statistically significant physics anomaly
- `|Z| > 5.0` → high-confidence synthetic indicator
- `|Z| > 9.0` → definitive — observed in commercial TTS systems

---

## Biological Baseline

Z-scores are computed against the SSA bifurcated VCTK/AMI baseline:

| Baseline | Corpus | N | Use |
|----------|--------|---|-----|
| `vctk_pool` | VCTK 0.92 anechoic (MKH800), pooled | 88,326 | Spectral, biomechanical, IAIF, voice quality features |
| `vctk_M/F` | VCTK, gender-stratified | M: 40,321 / F: 47,833 | Formants, pitch features |
| `ami_M/F` | AMI Meeting Corpus, gender-stratified | M: 21,779 / F: 8,850 | Macro-prosody (pause, F0 declination, nVIV) |

Gender is inferred from pitch_mean via the SSA 165 Hz F0 pivot threshold.
MLAAD contains no speaker gender labels.

---

## Quick Load

```python
import pandas as pd

df = pd.read_parquet('mlaad_with_zscores_20260522.parquet')

print(f"Rows: {len(df):,}  Columns: {len(df.columns)}")
print(f"Languages: {df['language'].nunique()}")
print(f"TTS systems: {df['tts_system'].nunique()}")

# Primary detection features by TTS system
df.groupby('tts_system')[['Z_bico_f0_f1', 'Z_spectral_tilt',
                           'Z_modgd_var', 'Z_nVIV']].mean().round(3)
```

```python
# nVIV rhythm collapse by language
# Expected: tonal/syllable-timed languages ~30-45
# Observed: uniform collapse toward ~56 regardless of language
df.groupby('language')['nVIV'].agg(['mean', 'std', 'count']).sort_values('mean')
```

```python
# High-quality clips only (Tier 1 = PRISTINE conditions)
tier1 = df[df['tier'] == '1']

# IAIF glottal features (most meaningful in Tier 1)
tier1[['tts_system', 'Z_iaif_residual_kurtosis',
       'Z_iaif_gci_regularity', 'Z_iaif_hf_energy_ratio']].groupby('tts_system').mean()
```

---

## Known Nulls

| Column(s) | Null count | Reason |
|-----------|-----------|--------|
| All raw features | 119 | Clips with extraction failures at ingestion |
| `spectral_aliasing_ratio`, `Z_spectral_aliasing_ratio` | 298,000 (all) | MLAAD is 16 kHz audio; 12-16 kHz band is above Nyquist ceiling. Expected and correct. |
| `jitter_local`, `shimmer_local`, `hnr_mean`, `cpps` | varies | PRISTINE gate — suppressed for T3/T4 clips (SNR < 50 dB or C50 < 50 dB) |
| `fam_75hz_sharpness`, `fam_86hz_sharpness`, `inertial_decay_residual` | 3,945 | Extraction threshold-dependent |
| `transcript`, `phonemes`, `vot_candidates` | 119 / 298,000 | Empty scaffolding columns from ingestion pipeline |

---

## Pipeline Summary

The SSA applies six sequential extraction passes to each clip:

| Pass | Tool | Features |
|------|------|---------|
| P1 | Brouhaha (Lavechin et al. 2022) | SNR, C50, speech ratio, quality tier |
| P2 | Parselmouth/Praat, librosa, scipy | 54 classical features |
| P3 | IAIF (Alku 1992), librosa LPC | 7 glottal residual features |
| P4 | Parselmouth + scipy VAD | 6 macro-prosody features |
| P5 | scipy, Parselmouth | 4 kitchen sink features |
| P6 | numpy/scipy (pure FFT) | 6 forensic features |

Full pipeline documentation: `docs/ssa_METHODOLOGY.txt`

---

## Key Finding: nVIV Rhythm Collapse

Across 100+ languages in MLAAD, TTS systems collapse nVIV (rhythm variability)
toward English stress-timed values (~55-65) regardless of the target language's
typological class. Tonal languages (Mandarin, Thai, Yoruba) and syllable-timed
languages (French, Spanish, Italian) would be expected to show nVIV ~30-45.
Observed across MLAAD: ~56.75 uniformly.

This is a generational finding — modern large-scale systems show improvement,
particularly in languages well-represented in training data. Use `Z_nVIV` as a
routing gate for language-typology analysis rather than a binary classifier.

---

## Distribution

**NON-COMMERCIAL — NOT FOR PUBLIC DISTRIBUTION**

This file is provided as a research audit contribution to the MLAAD dataset
team. It is not a general public release.

Contact: chris@moonscapesoftware.ca

---

## Citation

If you use these measurements in your research, please cite:

**SSA pipeline:**
```
Moonscape Software (2026). Synthetic Speech Atlas: Physics-Grounded Acoustic
Feature Extraction for Speech Deepfake Detection. SSA_MLAAD_v1 audit export.
Contact: chris@moonscapesoftware.ca
```

**MLAAD dataset:**
```
Müller, N. et al. (2024). MLAAD: The Multilingual Audio Anti-Spoofing Dataset.
Proceedings of Interspeech 2024.
```

**Brouhaha:**
```
Lavechin, M. et al. (2022). Brouhaha: Mono-channel speech assessment toolkit.
Interspeech 2022.
```

---

*Moonscape Software | Synthetic Speech Atlas | SSA_MLAAD_v1*
*chris@moonscapesoftware.ca*